🤖 AI

Agent-Omit: Adaptive Context Omission for Efficient LLM Agents

This paper introduces Agent-Omit, a unified training framework that enhances LLM agent efficiency by enabling them to adaptively omit redundant thoughts and observations through cold-start fine-tuning and omit-aware reinforcement learning, achieving superior effectiveness-efficiency trade-offs across multiple benchmarks.

Original authors: Yansong Ning, Jun Fang, Naiqiang Tan, Hao Liu

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yansong Ning, Jun Fang, Naiqiang Tan, Hao Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Over-Explainer" Agent

Imagine you hire a very smart, but slightly obsessive, research assistant to find a specific fact on the internet for you.

The Good News: They are brilliant. They find the answer.
The Bad News: To get there, they write a 50-page diary entry for every single step they take. They write down, "I am thinking about clicking the blue button," then they click it. Then they write, "I am reading the text on the screen," then they read it. Even when they are just clicking a button that was obvious from the previous step, they write a whole paragraph about it.

This is exactly what current AI agents do. They generate a lot of "thoughts" (reasoning) and "observations" (reading the results of their actions). While this helps them solve hard problems, it creates a massive amount of "digital clutter." This clutter:

Costs a lot of money (because AI companies charge by the word/token).
Slows everything down.
Confuses the AI because the conversation history gets so long it forgets the beginning.

The Insight: Not All Thoughts Are Created Equal

The researchers behind this paper asked a simple question: "Do we really need to read every single page of that 50-page diary?"

They analyzed the AI's journey and found that the importance of these thoughts and observations changes depending on when they happen:

The Beginning: The AI needs to think hard to plan its route. (High value).
The Middle: The AI is just clicking buttons or reading simple search results. Often, it doesn't need to write a diary entry for this; it can just do the action. (Low value).
The End: The AI needs to summarize the findings. It needs the final results, but it doesn't need to re-read the search results from three hours ago. (Low value for old data).

The Analogy: Imagine you are following a recipe. You need to read the instructions carefully at the start (planning). But when you are just stirring the pot for the 10th minute, you don't need to write a novel about how you are stirring. You just stir. If you keep writing about stirring, you run out of paper before you finish cooking.

The Solution: Agent-Omit

The team built a new training framework called Agent-Omit. Think of this as teaching the AI a new superpower: The Art of Knowing What to Skip.

Instead of forcing the AI to write everything down, Agent-Omit teaches it to say, "I know what I'm doing, I don't need to write this down," or "I don't need to re-read that old note."

They did this in two steps:

Step 1: The "Cold Start" (Teaching the Basics)

First, they created a special set of practice problems. They manually showed the AI examples of:

"Here is a time you thought too much. Next time, just skip the thought."
"Here is a time you read an old observation that didn't help. Next time, skip reading it."

They taught the AI the specific "hand signals" (special codes) to use when it wants to skip a step. It's like teaching a student that sometimes, "I know the answer" is a valid response, and they don't need to show their work for every single math problem.

Step 2: The "Reward System" (Reinforcement Learning)

Next, they let the AI play games (like navigating a website or solving a science puzzle) and gave it a special reward system:

Reward for getting the answer right.
Bonus Reward for using fewer words.

If the AI tried to be lazy and skipped a necessary step, it got a bad score. But if it correctly identified a redundant step and skipped it, saving money and time, it got a huge bonus. Over time, the AI learned to be a "smart minimalist"—doing the work but writing less about it.

The Results: The "Goldilocks" Agent

The researchers tested this new AI (called Agent-Omit) against seven other top-tier AI agents.

Performance: It solved the problems just as well as the "super-smart" giants (like DeepSeek-R1 or o3).
Efficiency: It used significantly fewer words (tokens) to get there.
The Trade-off: It found the perfect balance. It wasn't so lazy that it failed, and it wasn't so chatty that it wasted money.

The Analogy: Imagine two drivers trying to get to the same destination.

Driver A (Old AI): Drives the whole way while narrating every turn, every traffic light, and every cloud they see. They get there, but they burn a lot of gas and take a long time.
Driver B (Agent-Omit): Drives the whole way, but only speaks up when they need to make a complex decision or check a map. They get there just as fast, but they use much less gas.

Why This Matters (According to the Paper)

The paper claims that by teaching AI to "omit" (skip) unnecessary context, we can make these agents:

Cheaper to run.
Faster.
Just as smart.

They proved mathematically that this skipping doesn't hurt the AI's ability to think correctly, as long as it skips the right things. They showed that the AI naturally learned to skip about 3 to 4 rounds of unnecessary chatter in the middle of a task, which is exactly where the "boring" parts of the work usually happen.

In short, Agent-Omit is a training method that teaches AI agents to stop over-explaining, saving time and money without losing their intelligence.

) or issue specific commands to omit historical observations (e.g., <omit tool response N ...>`).
* Multi-Turn Omission: Constructs full trajectories where identified redundant turns are replaced with omission behaviors, forcing the agent to maintain reasoning continuity despite missing context.

Fine-Tuning: The agent is fine-tuned on this synthetic data using standard language modeling loss with a mask applied to environmental observations.

Stage 2: Omit-Aware Agentic Reinforcement Learning
To generalize beyond synthetic data and handle dynamic interactions, the authors introduce an RL framework featuring:

Dual Sampling Strategy: To address the "context change" problem (where an agent cannot learn from a decision it made after an omission), the training decouples sampling into:
- Full Trajectory (y): Evaluates overall efficiency and final task success.
- Partial Trajectory (y'): Samples the context before an omission occurs, allowing the agent to learn the omission policy conditioned on the pre-omission state.
Omit-Aware Reward: A composite reward function balancing task correctness ( $R_{task}$ ) and token efficiency ( $R_{omit}$ ). The omission reward is calculated based on the ratio of saved tokens but is set to zero if the task reward is zero, preventing reward hacking.
Policy Optimization: The agent policy is optimized using Group Relative Policy Optimization (GRPO) to maximize expected rewards across both trajectory types while minimizing KL-divergence from a reference policy.

2.3 Theoretical Guarantee

The authors provide a theoretical proof that the deviation of the learned omission policy from the optimal policy is upper-bounded by the KL-divergence between them. Assuming semantic Lipschitz continuity of task reward and token cost functions, the framework guarantees that as the policy approximates the optimal omission strategy, the deviation in effectiveness and efficiency converges.

3. Key Contributions

Unified Analysis Framework: The paper establishes a turn-level analysis framework quantitatively proving that omission mechanisms can improve agent efficiency without sacrificing effectiveness.
Agent-Omit Framework: A novel training paradigm integrating omission behavior synthesis (cold-start) with omit-aware agentic RL to train agents capable of adaptive context management.
Empirical and Theoretical Validation: Extensive experiments across five benchmarks and theoretical analysis demonstrating the approach's effectiveness in achieving superior effectiveness-efficiency trade-offs.

4. Experimental Results

The framework was evaluated on five diverse benchmarks: DeepSearch, WebShop, TextCraft, BabyAI, and SciWorld.

Performance vs. Frontier Models: The Agent-Omit-8B model achieved accuracy comparable to seven frontier LLM agents (including DeepSeek-R1-0528, o3, and Qwen3-32B) while substantially reducing token costs.
Efficiency vs. Existing Methods: When applied to Qwen3-8B, Agent-Omit consistently outperformed seven existing efficient agent construction methods (including Thought Management, Observation Management, and Summarization-based approaches). It achieved the highest Pass@1 accuracy with the lowest average token cost across all benchmarks.
Ablation Studies: Results confirmed that both the SFT cold-start (specifically single-turn omission data) and the RL stage (specifically partial trajectory sampling and omission rewards) are critical for performance.
Behavioral Analysis: Trained agents adaptively omitted an average of 3–4 turns of redundant context per trajectory, with omissions predominantly occurring in intermediate turns, aligning with the initial quantitative findings.

5. Significance and Claims

The paper claims that Agent-Omit offers a new paradigm for efficient LLM agents by moving away from static, heuristic-based context compression toward adaptive, learned omission policies.

Efficiency: The approach significantly reduces token consumption and computational costs, which the authors note may contribute to lowering the energy footprint of deploying large language model agents.
Scalability: By enabling smaller models (e.g., 8B parameters) to match the performance of much larger frontier models while using fewer resources, the method enhances the practical applicability of LLM agents in resource-constrained environments.
Limitations: The authors modestly acknowledge that the current approach incurs additional training samples during the RL stage, which limits scalability. Future work is proposed to scale the omission data synthesis pipeline to large-scale pre-training and extend the paradigm to larger LLMs.

The authors declare no potential financial conflicts of interest related to the work.